Reward-based e-commerce companies expose thousands of online offers and coupons every day. Customers who signed up for online coupon services either receive a daily digest email with selected offers or select specific offers on the company website front-page. High-quality online discounts are selected and delivered through these two means by applying a manual process that involves a team of experts who are responsible for evaluating recency, product popularity, retailer trends, and other business-related criteria. Such a process is costly, time-consuming, and not customized on users’ preferences or shopping history. In this work, we propose a contentbased recommender system that streamlines the coupon selection process and personalizes the recommendation to improve the clickthrough rate and, ultimately, the conversion rates. When compared to the popularity-based baseline, our content-based recommender system improves F-measures from 0.21 to 0.85 and increases the estimated click-through rate from 1.20% to 7.80%. The experimental system is currently scheduled for A/B testing with real customers.
Reward-based e-commerce services are a fast-growing online sector that provides cashback to subscribers by leveraging discounted prices from a broad network of aliated companies. Typically, a consumer would use the company website to search for specific products, retailers or potentially click on one of the prominent published offers. Shoppers are then redirected to the websites of the respective retailers. A visit to the merchant’s website generated by this redirection is dened as shopping trip. At the same time, subscribers receive daily digest email with the most popular discounts. Figure 1 shows a fragment of a daily digest email featuring a list of current offers and coupons. Both website and email contents are manually curated by experts focusing on delivering the highest quality offers to customers. This manual process is laborious and does not scale well to millions of online customers who are receiving the same oer recommendations regardless of their previous browsing or purchase history. A recommender system would be able to capture the customers’ preferences by optimizing the coupon selection based on the previous history and the similarity across users and / or items. However, compared to traditional recommender systems in other domains, such as movies [6, 10] and music , online offers dier for a number of reasons: 1) Coupons are highly volatile items only valid in a limited time span and removed after their expiration date; 2) Customers using coupons are aected by high turnover; 3) Users’ population is skewed between a minority of heavy users and a majority of occasional or light users; and nally, 4) Online offers are rarely rated like usually happens for movie and music domains. Consequently, the online discount domain is more prone to the cold-start problem , where there is insucient click history for a new item or, conversely, there is not enough history for a new subscriber. Matrix factorization methods are less robust to the coldstart problem  and lack the capability to capture the number of features necessary to model preferences of high volatile items and large customer populations. For these reasons, we adopted a content-based ltering (CBF) approach  where we exploit coupons and customers’ attributes to build a stochastic classication model that predicts the posterior probability for a coupon to be clicked by the customer. We rely on users’ implicit feedback which indirectly express users’ preferences through behavior like clicking on specific offers in a ranked list. This is typically a more noisy signal compared to explicit feedback such as ratings and reviews, but it is also abundant and with consistent bias across user’s click-through history [15, 23]. However, click information is more challenging to use since it does not directly refect the users’ satisfaction neither provides clues about items that are irrelevant (negative feedback)  which is fundamental to build an eective discriminative model. In this paper, we illustrate how to utilize noisy implicit feedback to build a CFB model for offers and coupons recommendations. The main contributions are the following: (1) We describe a method to generate negative samples from the implicit feedback data; (2) We perform extensive experiments to compare dierent learning models; (3) We demonstrate that our approach is eective with respect to the baseline and the upper-bound baseline.